Space Bunny Was MiniMax M3.1 Flash

The one-sentence version

Space Bunny 🐰, which processed more than 200 trillion tokens while operating anonymously, was an early version of MiniMax M3.1 Flash. That sentence is MiniMax’s own, from the launch post on 19 October 2026.

The scale is the part worth pausing on. 200 trillion tokens is not a demo, a beta or a weekend experiment — it is sustained production traffic through a model that had no name, no price, no benchmarks and no vendor attached to it.

The timeline

Everything below is UTC and dated. The reveal came 25.8 days after the listing — the longest anonymous run of any stealth model on OpenRouter.

The reveal date is the only one MiniMax chose. Everything before it was inference by the community.

How an anonymous model gets identified

The method is called tokenizer fingerprinting, and it is the reason the mystery was solved in a quarter of an hour.

A tokenizer is the model’s tokenizer pair: the step that turns your text into tokens. Different labs use different tokenizers, and a tokenizer’s behaviour on unusual strings is distinctive. So you send the endpoint a set of test strings, read back the input-token count the API reports, and compare that against the public tokenizers of known model families. A match rules families in or out.

What it does and does not prove is the key caveat:

• A match rules a family in or out. Qwen, GLM, DeepSeek and Llama tokenizers were all tried and did not match.
• A match rarely names the exact release. Models inside one family often share a tokenizer, so the honest conclusion was “MiniMax family” — one tester wrote exactly that, adding “M2.7 matched too, so this points to the family rather than a specific release.”
• The same technique had already been used successfully on OpenRouter’s Ox Alpha, which was fingerprinted as GLM from its tokenizer and confirmed when Z.ai named it six days later.

The published traits lined up too

By 27 September the public MiniMax Code config listed M3.1-Flash-Preview with three traits, and all three matched the OpenRouter listing for Space Bunny Alpha:

Three matching traits is a strong hint, not a name — and MiniMax still did not say anything for another 22 days.

OpenRouter’s stealth model record

Space Bunny was not the first, and it was by a wide margin the longest.

Every OpenRouter stealth launch before Space Bunny ended with the maker being named and the free price ending with it. Union Alpha is the interesting counter-example: it fooled the fingerprinters because it was a blended model from a startup nobody had guessed. Space Bunny looked like the Ox Alpha case instead — a clean tokenizer match to one lab’s family, and then a same-spec product from that lab.

What the anonymous window actually measured

While the model was running under a name nobody could claim, OpenRouter was still publishing endpoint-level numbers for it. These are third-party measurements of the stealth endpoint, and they are the only hard data that existed before the launch.

Chinese-language coverage of the launch reported larger cumulative figures from the same period: Space Bunny passing 56.7 trillion tokens on OpenRouter over 10 consecutive days at number one, 106 trillion tokens on OpenCode over 12 consecutive days at number one, and taking the following week’s chart with 38.7 trillion tokens — 13 trillion ahead of DeepSeek V4.1 Flash. These are third-party reports, not numbers MiniMax published.

What happened to your prompts

The data terms during the anonymous window were better than most people assumed, and still worth reading carefully.

OpenRouter’s launch notes said: “This time, the provider does not train on your prompts or completions.” The listing was more precise: “Prompts and completions may be retained by the provider but are not used for training”, with everything else under OpenRouter’s Stealth Model Terms.

No training is a real promise. It is not a promise of deletion, and the company holding the data was not named — the provider behind a stealth endpoint is exactly the thing you cannot name. OpenCode described its own route as keeping no data, which is a claim about OpenCode’s route rather than the OpenRouter endpoint.

The practical rule: for code or documents you would not hand to an unknown company, wait for the name.

What the reveal actually means

• Engineering maturity. A model that survives 200 trillion tokens of production traffic is not a research artefact. Stability, throughput and cost behaviour were proven before anyone could name it.
• The cost story is the differentiator. If M3.1 Flash is an evolution of something that ran at that scale, the honest reading of the “70% cheaper agentic task cost” claim is that it describes a production economics result, not a marketing estimate.
• It is not a capability ceiling. Space Bunny was an early version. The public model is a later revision. Inferring today’s capability from anonymous traffic would be a mistake.
• The community had already stress-tested it. For three and a half weeks, strangers ran it hard with no documentation and no support, and the model built its reputation on what it actually did. That evaluation is written up separately on the Space Bunny evals page.

Is Space Bunny still around?

At the time of writing, the public model is M3.1 Flash and the Space Bunny label belongs to history. MiniMax’s launch post says the model weights “will be released later”, and the MiniMax organisation on Hugging Face has no M3.1 repository yet — the most recent additions are MiniMax-M3 (July), MiniMax-H3 and MiniMax-Music3 (August). We have not found an authoritative statement about whether the anonymous endpoint still exists or what it now serves. If you find one, tell us and we will cite it here rather than guess.

Next: the community’s hands-on evaluation of the anonymous model is on the Space Bunny evals page; what the capabilities mean in practice is on the use cases page; the measured numbers are on benchmarks.